1) order flow data is not useless because it comes from uninformed traders. It depends on the model being used. Most hft models aren't even concerned about the inflow of big trades from institutional money. Most of their models are simple short period time series forecasts they trade around. Which means the model is pretty agnostic on where the trades come from as long as enough back tests show that their predictive power is good enough.

2) even if that were not true, hft works on volumes. So I contradict my previous speculative comment by saying this but volumes of institutional flows absolutely dwarves retail flows. You'd have much more opportunities trading around institutional money than retail money.

Surely the order flow is mostly going to look like noise, random. But things like local information and employee information would provide some signal - correlated trading patterns - that indicates there could be some information available to inform a trade. Allowing one to either trade on the meta-signal or seek the information behind it (and then trade).

So if you've got a company and the city in which it's headquartered just gets a strong buy signal, sure that could be random but I'd imagine ...

Surely having the meta-data on who is trading and linking that to trades is the primary benefit?